Firm AI-Spend Intensity and Headcount Growth

Sources#

• 2026 State of Scaling: The Great Sorting
• A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment
• AI and Job Postings: From Destruction to Creation?
• AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It
• Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think
• Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence
• China, Open Source & AI Competitiveness — Andrew Ng
• How Does AI Change Labor Demand? Evidence from 41 Countries
• Ramp's latest data on China vs. the American AI Labs
• September 2026 Ramp AI Index: Cracks in the AI thesis, part 2
• The Human-AI Substitution Principle: When will you be replaced by AI in your organization?
• The state of AI in 2026: On the road to ROI

Summary#

The headline result of Kharazian, Simon & Stevens (Ramp × Revelio Labs, June 2026): firms that adopt generative AI grow faster after adoption, but the effect is gated by intensity. Linking Ramp's line-item corporate-card and bill-pay records (which reveal actual payments to AI vendors) to Revelio Labs' workforce histories for 21,559 US firms, they find high-intensity adopters grow total headcount ~10.2% and entry-level headcount ~12.0% over the first 24 months after adoption, while low-intensity adopters show no statistically detectable change. Gains emerge gradually (a "learning curve"), are broad across job functions, but are concentrated in the Information sector — and the adopter population is heavily self-selected. The results "counter predictions that AI adoption will lead to broad job loss," at least at the firms doing the adopting.

The paper's methodological contribution is as important as its finding: it is, to the authors' knowledge, the first to combine observed firm-level AI spending with workforce records at scale, replacing occupational-exposure proxies and executive surveys with a revealed-adoption measure grounded in real spend. See Telemetry vs. Survey Measurement and Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated for where this instrument sits among the alternatives.

Evidence note. empirical — a staggered difference-in-differences design on administrative-grade spend + workforce panels, not a lab experiment. Two caveats travel with every claim below. (1) Selection: adopters are larger, more technical, higher-paying, faster-growing, and far more VC-backed before adoption (Table 2), so a naïve adopter-vs-never-adopter comparison confounds AI with pre-existing growth. (2) Identification: the preferred estimates rest on a conditional-parallel-trends assumption (Callaway–Sant'Anna, comparing adopters to not-yet-adopters in the same eventual intensity group with sector fixed effects). Pre-trends are clean for most outcomes but not all — high-intensity total headcount carries 3 of 11 flagged pre-periods, and Bachelor's/MBA carry 4–5. The mechanism (why adopters grow) is explicitly unresolved.

The novel instrument: spend-side adoption#

Prior work measured AI's labor effect through occupational exposure (which tasks an LLM could do — Eloundou et al.; Webb; Felten et al.) or surveys, because direct firm-level adoption data were unavailable — generative AI is bought through software subscriptions and API calls, not the trackable capital equipment earlier technology studies could benchmark. Exposure indices vary only across occupations, not across firms, so "two firms employing identical workers may differ sharply in AI adoption, and exposure indices cannot separate them." Even the best observed-exposure work (Massenkoff & McCrory's Anthropic-Economic-Index measure — the "observed exposure" of Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated) remains occupation-level.

Ramp's data closes that gap. Using the Ramp AI Index vendor/line-item classifier (foundational LLMs, GPU cloud, model serving/inference, coding agents, API tokens, AI image/video, AI search), the authors observe when a firm starts paying AI vendors, how much, and to whom:

• Adoption (G_i) = the first month of the earliest three-consecutive-month spell with ≥$100/month AI spend; absorbing thereafter. This excludes one-off employee experiments while capturing sustained organization-level purchases.
• Intensity = PEPM (per-employee-per-month AI spend over the first three post-adoption months, divided by baseline headcount). Low = bottom two terciles, High = top tercile. The gap is an order of magnitude: Low adopters average $2.78/employee/month, High adopters $33.67 — the difference between enterprise chat subscriptions and sustained investment in coding agents, APIs, and multiple models.

As an adoption rate, the linked panel's paid-adoption series sits between the firm-weighted Census BTOS benchmark (18% of firms / 32% employment-weighted) and executive surveys (Yotzov et al. ~69%; Atlanta Fed ~78%) — a "revealed paid-use" measure for a business-spend-active, tech-skewed population, not a nationally representative rate.

The intensity-gated result#

Preferred estimates (Callaway–Sant'Anna, not-yet-treated within intensity group, NAICS sector FE; log points ≈ percent, averaged over months 0–24):

Outcome	Low vs. Not-yet	High vs. Not-yet
Total headcount	−0.6% (ns)	+10.2%*
Entry-level headcount	−1.7% (ns)	+12.0%*
Non-entry headcount	+0.4% (ns)	+7.4%*
Manager-plus headcount	+1.0% (ns)	+6.5%*

(*** = p < 0.01.) The pattern across intensity groups is the paper's core evidentiary move: the firms spending the most on AI are the firms with the largest employment gains, while low-intensity adopters look like non-adopters. "Enterprise chat subscriptions do not appear to be enough… nor are a few months of experimental spending."

Gains compound on a learning curve. The high-intensity total-headcount event-study coefficient is ~0 at adoption (0.003), 0.020 at 3 months, then 0.071 (6mo) → 0.188 (12mo) → 0.277 (~32%, 18mo) → 0.452 (~57%, 24mo). The 24-month average of +10.2% therefore understates the endpoint; the late-window estimates carry wider confidence intervals as fewer firms are observed that far out. The earliest growth appears ~6–12 months after adoption — consistent with firms needing time to establish best practices, integrate tools, and then hire.

The entry-level result and the tension it creates#

The +12.0% entry-level figure (seniority 1–2 in Revelio's scale) is the paper's most narratively loaded finding: it runs directly against the widespread "AI is killing entry-level / junior jobs" claim. The clearest external statement of that claim — Brynjolfsson, Chandar & Chen, "Canaries in the coal mine" (Stanford Digital Economy Lab; cited in the raw doc's §2) — reports an ~16% employment decline for workers aged 22–25 in the highest-exposure occupations after ChatGPT's release. The Ramp result points the other way. (Superseded 2026-09-22: the vault now holds Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence directly. The current headline is a 19% kept-pace shortfall through June 2026, on a different measure from the 16% — see the reconciliation below.)

These do not cleanly contradict; they measure different objects (flag with tiers — both empirical). Brynjolfsson et al. identify off cross-occupation exposure within firms (young workers in exposed occupations decline economy-wide). Kharazian et al. identify off cross-firm adoption (at firms that adopt AI intensively, entry-level grows). Both can hold at once if junior-role contraction concentrates in exposed occupations at non-adopting or low-intensity firms — or reflects reallocation toward the adopting firms — while intensive adopters expand entry-level headcount broadly. Neither directly refutes the other; they are complementary margins (occupation-exposure vs. firm-adoption), and the unit of analysis is the whole difference. The Ramp paper is objective counter-evidence to the entry-level-job-loss fear that surfaces as perception in The Automation–Optimism Link (over ⅓ of surveyed workers put a junior colleague's job-loss probability above 60%) — but only for the adopting-firm population it observes.

The counter-instrument, read firsthand and reconciled rather than averaged (2026-09-22). The vault now holds the Canaries paper itself — Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, revised August 2026, ADP administrative payroll microdata through June 2026 (empirical; instrument caveats on Stanford Digital Economy Lab). Three corrections and one sharpening:

• The number moved, and on a changed measure. The paper de-emphasized the firm-shock-adjusted regression it used to headline (13% at the July 2025 vintage, 16% at September 2025) in favour of a plain descriptive divergence that "requires no modeling choices": 15% at the July 2025 vintage, 19% as of June 2026. So 16% and 19% are not two points on one widening series; the vault should stop quoting them as if they were. In levels, 22–25-year-olds in the two most-exposed quintiles fell ~11% Nov 2022 → Jun 2026 while the three least-exposed quintiles grew ~10%.
• The tension does not resolve — and the paper explains structurally why it cannot. ADP firm identifiers are anonymized, which "precludes merging external adoption measures based on job postings or earnings calls." Canaries therefore cannot see firms at all; it identifies off cross-occupation exposure, this panel off cross-firm adoption. Both empirical, both can hold: +12.0% entry-level at intensively adopting firms is compatible with −19% for young workers in exposed occupations economy-wide if junior contraction concentrates in exposed occupations at non-adopters, or if adopting firms are absorbing reallocated juniors. Neither instrument can test that reconciliation, and averaging them would produce a number describing nothing. (A third instrument tested it on 2026-10-01: Chandar & Klein Teeselink uses the same Revelio seniority 1–2 coding as this panel, with postings-defined adoption, a cross-border peer instrument, and 41 countries. It finds the junior share at adopters falling 1.9pp against matched non-adopters, junior employment −2.5% (n.s.), and senior employment +6.7%. That contradicts the "adopters absorb reallocated juniors" branch. Its own uninstrumented matched design reproduces this panel's direction (junior +7.0%, total +10.8%) and fails its pre-trend test. Comparison table and weighting on Seniority-Biased AI Adoption: The Junior Share at Adopting Firms.)
• The nearest thing to a bridge, and it leans Canaries' way. Lacking firm identifiers, the paper substitutes state-level adoption tiers built from Anthropic Economic Index usage: the decline for exposed young workers is sharpest in leading-adoption states (about −19% for the most-exposed quintile) and smallest in emerging-adoption states. That is an adoption gradient in the opposite direction from this panel's — geography is a crude proxy for firm adoption, but it is the only overlapping cut the two instruments have, and it does not corroborate the "adopters hire juniors" reading.
• A caution that cuts the other way, in Canaries' own appendix. The estimate that best controls for the confound this panel is built to address — within-firm Poisson event studies with firm-time fixed effects, i.e. "were these young people just at shrinking firms?" — attenuated across vintages and is no longer significant for the most exposed quintile (−11.7 log points, p=0.02 in the August 2025 release; −5.3, p=0.26 applying the same filters to current data). The fourth quintile is stable. The descriptive divergence widened while the firm-conditioned estimate weakened, and both facts belong in any citation.

Composition, not just scale. In workforce shares (Table 4), high-intensity adopters tilt younger: entry-level share +1.15pp, manager-plus share −1.52pp — entry-level headcount grows faster than the rest of the firm. Low-intensity adopters move the opposite way (entry-level share −0.52pp, manager-plus +0.66pp). So the intensive-adoption story is not merely "bigger firms" but "bigger and more junior-weighted."

The demand-side companion: job postings, and where they disagree#

Everything above is a headcount stock at firms observed paying AI vendors. Indeed Hiring Lab (Guillermo Gallacher, July 8 2026) supplies the flow side of the same question — vacancies posted, economy-wide rather than adopter-only. It points the same direction on level and the opposite direction on composition.

What the post states in prose (every figure below appears in its body text as well as a chart caption):

• US software-development postings +15% since Claude Code's launch (the series is indexed to 100 at February 24 2025), while overall postings fell 7% over the same window.
• The rebound starts from a deep hole: software-development postings remain ~27.5% below their February-2020 level, while overall postings are essentially back to it.
• The rebound is concentrated, not broad: 71% of the May 2025 → May 2026 increase in software-development postings comes from senior roles and 37% from postings whose title mentions AI (the two overlap). Gallacher's reading: "demand is growing for experienced professionals who can work with AI, not necessarily a broad-based recovery across all software roles."
• Across occupations the exposure–postings relationship flips sign by window: over May 2022 → May 2026 the more AI-exposed an occupation, the more its postings fell; over May 2025 → May 2026 the more exposed, the more they rebounded. Both are described as statistically significant and explicitly uncontrolled.

Evidence note and conflict of interest. empirical in that it is measured administrative data, but it is Indeed's Hiring Lab analyzing Indeed's own job board, published as a blog post rather than a study. Three limits travel with every number. (1) Postings ≠ jobs. One job board's vacancy flow is a selected slice of labor demand — skewed by which employers and sectors post there, by country coverage, and by posting behavior (reposts, evergreen listings) that a headcount panel does not have. (2) The causal attribution is asserted, not identified. The post says outright that "correlation does not imply causation," then leans on the coincidence anyway: Claude Code's launch, the coining of "vibecoding," and the postings trough all sit in February 2025. A rebound that begins at a product launch is not evidence the product caused it, and the post's own citation of the New York Fed notes the AI-exposed decline began before ChatGPT's release, so the trough is the end of a pre-existing trend as much as the start of a new one. Establishing causation needs variation this design has none of — differential exposure to agentic tooling across firms or geographies, a matched control set of comparable unexposed occupations, or adoption-timing variation of the kind the Ramp panel above actually exploits. Read it as a description of a turning point, not an estimate of an effect. (3) Chart-only quantities are deliberately not quoted here. The identities and per-country shares of the six economies in the international comparison, and both scatter plots' correlation coefficients and p-values, exist only inside static images hosted on the source site; there are no local copies to verify against. The prose says only that the software share of postings is rising in most large developed economies "with the exception of Germany and France," and that the exposure relationships are "statistically significant." (Minor source-internal inconsistency: the international chart's own caption names only Germany as declining, against the body text's Germany and France. Neither is checkable without the chart, so the vault carries the prose version and this note.)

Where it agrees with the spend panel. Both instruments say AI-intensive labor demand is rising rather than collapsing, and both find it concentrated in software rather than economy-wide (Ramp: significant only in Information; Indeed: software development against a declining overall market). Indeed's sign flip is also the vacancy-side echo of this page's core point — the same occupation-level exposure score predicts decline in one window and growth in the next, so exposure is not a fixed verdict on a job.

Where it disagrees, and it is on the sharpest number. Ramp's compositional headline is that intensive adopters tilt junior: entry-level headcount +12.0%, the fastest band, with entry-level share +1.15pp against manager-plus −1.52pp. Indeed's compositional headline is that the postings rebound tilts senior: 71% of the increase. These are not descriptions of the same object:

	Ramp × Revelio	Indeed Hiring Lab
Unit	firm-month headcount (stock)	job-board postings (flow of vacancies)
Population	21,559 US firms that pay AI vendors	all US postings on one job board
Contrast	adopters vs. not-yet-adopters, sector FE	before vs. after Feb 2025, no control
Window	months 0–24 post-adoption	May 2025 – May 2026
Composition	entry-level fastest (+12.0%)	71% of the increase is senior

They are reconcilable in principle — intensive adopters can expand junior headcount while the marginal vacancy posted economy-wide skews senior, and a posting flow says nothing about the seniority mix a firm already employs. But that reconciliation is a hypothesis, not a measurement, and the direction matters: Indeed's split lines up with the Brynjolfsson "Canaries" pattern (a 19% kept-pace shortfall for 22–25-year-olds in exposed occupations through June 2026, against no comparable gap at any older age — the −16% / +8% figures previously quoted here were Lovett's secondhand rendering of an older vintage) that the Ramp entry-level result cuts against. Gallacher concedes the point himself — even with the rebound, "the job market could still be experiencing a seniority-biased technological change." The vault now holds one instrument on each side of the junior question, on different units, and the disagreement is unresolved.

A third voice, with no instrument at all. Andrew Ng, interviewed by the Washington Post in July 2026 (China, Open Source & AI Competitiveness — Andrew Ng, practitioner-opinion), asserts the same direction from the practitioner side and generalizes it into a forecast: "software engineering job postings are up… and the industry is healthy and growing," against what he calls the "fear narrative" of a job apocalypse — "some people have said 50% of the people [will lose their] job, [there] may be rioting in the streets. That's not going to happen." His mechanism is that AI raises the value of the complementary human: "frankly we can't get enough skilled AI engineers… AI has made software engineers even more valuable." He then extends software to the rest of the economy — "software will prove to be a harbinger or a forerunner of a trend we'll see in other sectors" — landing on reskilling rather than displacement as the binding problem: "the challenge is not dealing with the job apocalypse… the challenge is how to help everyone gain the new skills they will need."

Worth recording for two reasons and worth discounting for one. It is a third independent arrival at rising-not-collapsing AI-adjacent labor demand, and it supplies a mechanism the two measured instruments do not (complementarity raising the price of the human, not merely adopters staffing up). But he offers no data for any of it, and his summary is precisely composition-blind in the way the table above exists to correct: "the industry is healthy and growing" is compatible with Indeed's +15% and with the same page's 71%-senior split and the ~27.5% hole below February 2020 that the rebound is climbing out of. The generalization to other sectors is a prediction, and the vault's own sector cut is the reason to hold it loosely — the measured headcount effect is significant only in Information, which is the sector Ng is generalizing from.

Restated a month later, with the condition made explicit (August 2026). In the Silicon Valley Girl interview (Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think, 2026-08-28, practitioner-opinion) Ng repeats the postings claim — "the number of job openings in software engineering is up contrary to what… the doom fear-mongerers would say" — and adds three unsourced observations: "all the good software engineers I know are busier than ever," employers he knows "can't find enough skilled… people at any level of seniority," and his own office's interns, including a high-school student, are "amazing and productive" because "they're all very AI native." The addition that matters is the condition he now attaches: "if someone still write code like… 2022 before ChatGPT they're in trouble… gain your skills to do the other 60 70% that AI cannot do." That is a composition claim — demand is up for one kind of engineer and down for another — which is precisely what the 71%-senior, 37%-AI-titled shape of Indeed's rebound in the table above measures, and what "the industry is healthy and growing" elided in July. His interns are the one datum here bearing on the entry-level split, and they are an n-of-one office run by the person making the argument.

Broad across roles, concentrated in one sector#

For high-intensity adopters, growth spans functions (Table 3): sales +10.3%, admin +7.8%, engineering +7.3%, entry-level engineering +6.3%, customer service +6.3%, scientist +5.6%, with finance (+4.6%) and marketing (+5.7%) marginally significant (marketing has weak pre-periods). Operations is the only category that did not grow. Notably, there is no strong correlation between a function's perceived AI exposure and whether its headcount rose or fell — cutting against the intuition that "exposed" functions shrink. Education outcomes echo this (Bachelor's +10.3%, MBA +4.9%) but with more flagged pre-periods; JD is unchanged and PhD is positive-but-insignificant.

Sector concentration is the sharpest limit on the headline. Estimated within broad NAICS groups (Table 5), the total-headcount gain is statistically significant only in Information (+13.4%, 0/11 pre-period flags) — the group of software, internet, and media firms. Professional/technical services is positive but insignificant; the rest hover near zero. The authors read this as "it is still early": AI's commercially mature gains are clearest in coding-agent and software-engineering workflows, where cheaper core output raises the return to expanding the whole firm, and diffusion to non-technical sectors requires workflow redesign AI vendors haven't yet nailed. This makes the "AI grows jobs" claim, as measured, a claim about the Information sector today — not yet a broad-economy fact.

Robustness against never-adopters#

Against never-treated firms (Table 6, descriptive not preferred), the estimates are larger and still ordered by intensity — total headcount +12.6% (High), +6.5% (Low) — but the pre-trends are badly contaminated (Low 11/11, High 8/11 flagged pre-periods), exactly the selection the not-yet-treated design exists to defuse. The authors report it "for completeness," which is the right weighting: it inflates the effect because adopters were already growing faster than permanent non-adopters.

The same instrument, published monthly: the Ramp AI Index (July 2026)#

The paper above is a research use of the Ramp AI Index vendor/line-item classifier. Ramp also ships that classifier as a monthly market index, and Ara Kharazian's July 8 2026 letter (empirical, Ramp's lead economist) is the vault's first entry from it. Same payment rail, different cut: not adoption→headcount, but who US businesses are paying, month by month, back to January 2023.

Read it as Ramp's own book. Ramp is measuring its own corporate-card and bill-pay customers — a business-spend-active, VC-forward-skewed base, not a random sample of US businesses (the paper's own framing of its adoption rate, above, says the same). Ramp also has a commercial interest in being the authoritative AI-adoption data source, so the index's headline framings are marketing surface as well as measurement. Directions of trends are far more defensible than levels. Two denominators travel with the numbers and are easy to conflate: the vendor-share series is a share of eligible businesses (54.96% of which had any AI spend in June 2026), while the model-serving and spend series are shares of AI-spending businesses.

The vendor race, with a date the prose does not give. The letter reports Anthropic at 42.4% of businesses in June 2026 against OpenAI 39.5% ("essentially flat, edging down 0.1 points"). The recovered monthly series says something sharper than "flat":

• The crossover is datable to May 2026. April 2026: OpenAI 39.63% vs Anthropic 38.57%. May: 39.54% vs 41.01%. Anthropic passed OpenAI in one month and extended to +2.9pp by June.
• OpenAI peaked in November 2025 at 41.36% and has fallen in every month since February 2026 (41.17 → 40.28 → 39.63 → 39.54 → 39.47). Month-over-month it is flat; against its own peak it is down ~1.9pp. This is the first sustained decline in OpenAI's business-adoption share in the 42-month series.
• Anthropic's move is very recent and very steep: 18.40% (Dec 2025) → 42.40% (Jun 2026), +24.0pp in six months, against +7.8pp over the whole preceding twelve. Overall AI adoption grew only +9.0pp over the same six months, so most of Anthropic's gain is penetration into businesses that were already paying for AI, not new adopters.
• Derived from the same table (not stated by Ramp): among AI-spending businesses, OpenAI's penetration fell 87.5% → 71.8% between January and June 2026 while Anthropic's rose 46.2% → 77.2%. Rebasing on AI spenders rather than all businesses roughly doubles the apparent swing, because the denominator was itself growing.

Four vendors the letter never mentions, all recovered from the chart datasets. Google has been flat for the entire series — 4.73% in January 2023, 6.38% in June 2026, never outside 4.6–6.4% — while overall AI adoption went 7.5% → 55.0%. Microsoft reaches only 1.73%. xAI sits at 3.03%, with a visible one-month doubling (1.72% → 3.30%, June→July 2025) and a plateau since. DeepSeek's direct line is the most interesting of the four and the most conspicuously absent from an article about Chinese models: it peaked at 0.23% during the January 2025 R1 news cycle, then declined for a year, and its 2026 recovery reaches only 0.29% — roughly 1/20th of the model-serving proxy and 1/146th of Anthropic.

That pattern is also the instrument's clearest aperture warning. The two vendors with by far the largest enterprise-contract businesses — Microsoft and Google — are the two smallest lines here. The likely explanation is the payment rail, not the market: enterprise agreements, cloud committed-spend drawdowns and negotiated invoices do not land on a corporate card the way an OpenAI or Anthropic subscription does. Read the vendor shares as share of businesses paying a vendor through a card/bill-pay rail, which is a good instrument for seat-and-API purchasing and a poor one for hyperscaler consumption. See Telemetry vs. Survey Measurement.

Where it joins this page: the index measures PEPM in the paper's own units. The paper's intensity split is per-employee-per-month AI spend — low adopters averaging $2.78, high adopters $33.67. The index's June 2026 medians are $10.59 for the typical AI-spending business and $248.41 for businesses using model-serving/inference platforms. So the model-serving cohort is not merely "high intensity" — at ~7× the high-intensity group's own mean, it is an extreme tail inside it, which is the cohort this page finds growing headcount ~10% (~12% entry-level). The comparison is indicative rather than exact: the paper's PEPM is a cohort mean over each firm's first three post-adoption months across the panel window, the index's is a June-2026 median. Two further things the recovered series adds: the multiple widened from 14.9× (June 2024) to 27.4× (January 2026) and has since compressed to 23.5×, because the median AI-spending business nearly doubled its per-employee spend in five months ($5.59 → $10.59, +89%); the "23×" gap is currently closing, not opening.

And the finding the whole letter is built around: Chinese/open-model adoption is additive, not substitutive. Full treatment at The Open-Weight Frontier Gap — in short, 5.8% of AI-spending businesses use model-serving platforms (Ramp's proxy for open-source and Chinese models), and 96.4% of those still pay OpenAI or Anthropic directly, at rates above the AI-spender base rate rather than below it.

The next edition, and the first month the index turns down (September 2026)#

Kharazian's September 9 2026 letter (empirical, same instrument, same COI, same five-charts-as-Datawrapper-iframes construction) is part 2 of a serial he calls Cracks in the AI thesis, and the first edition written as a bear case on the model companies' revenue. Four of its series bear on this page.

Adoption is still growing and the growth decelerates every month. August 2026: 56.1% of eligible US businesses had AI spend (the letter's "56%, up 0.4 points"), against 55.0% in June 2026 (superseded 2026-09-22 by September 2026 Ramp AI Index: Cracks in the AI thesis, part 2); Anthropic 43.8% (+0.34pp MoM) against OpenAI 39.8% (+0.09pp), so the +2.9pp lead recorded above is now +4.0pp. The deceleration is the sharper reading and is visible only in the recovered series (vault arithmetic, not stated by Ramp): the monthly increment in overall adoption has fallen in every month since March — +2.17 → +1.71 → +1.19 → +0.78 → +0.76 → +0.43pp. Anthropic's own surge decays on the same schedule — +6.66 → +4.43 → +2.45 → +1.38 → +1.06 → +0.34pp — so the process that produced the May crossover is running out of firms to convert, and Anthropic's August gain is now smaller than xAI's.

Two vendors moved more than the headline pair, and the letter names neither — the same blind spot as the July edition. xAI added +0.66pp in August (3.97% → 4.63%), the largest single-vendor gain of the month and nearly twice Anthropic's; Microsoft reached 2.01%, from 1.35% a year earlier. Google is still flat at 6.15%, inside the 4.6–6.4% band it has never left in 44 months. DeepSeek's direct line makes a series high of 0.34% — still ~1/128th of Anthropic, and still unmentioned in a serial whose thesis is cheap models.

A new cut: per-employee spend by cohort, and the first decline in the series. This edition replaces the typical-vs-model-serving spend split with a three-cohort PEPM series. August 2026: median AI-spending business $12.50, top decile $675.46, top percentile $7,205.13. The top 1% is what the letter is about — down 9.7% month over month from an as-reported July value of $7,976, and it is the cohort that "drives the vast majority of enterprise revenues for the model companies."

The decline is confined to the one cohort that can least support it. The median AI spender went $10.94 (June) → $11.90 → $12.50 (+14% in two months) and the top decile $521.96 → $675.46 (+29%); only the top percentile fell. Ramp's own methodological note says that cohort "represents a small group of firms" and is "more volatile than our median and top 10% estimates," and Kharazian offers the seasonal reading himself — August engineer holidays, with comparable dips each November–December — before arguing past it.

And the vintage problem, which cuts the headline roughly in four. (Vault arithmetic across two editions of the same index; Ramp states the inputs, not this conclusion.) Ramp revises recent months upward as late transactions land, and discloses one such revision here: July top-1% PEPM went from ~$7.4K as first reported to ~$8.0K (+8%). Comparing the two ingested editions on their overlapping months shows the same thing in the other spend series — June 2026 median PEPM reads $10.59 in the July edition and $10.94 in the September one (+3.3%), and June model-serving adoption 5.77% → 6.00% (+3.9% relative) — while the binary vendor-share series barely moves (June overall adoption 54.96% → 54.95%). Count-of-firms series are near-final on first print; dollar series are not. The headline −9.7% therefore compares a revised July against an unrevised August. Like-for-like at first print, $7,205 against ~$7.4K is about −2.6%, and August will most likely revise up as well. The decline may be real; its published magnitude is partly a vintage artifact — and every month-over-month change read off the latest point of this index is biased downward for the same reason.

Where it joins this page. The index publishes this panel's intensity variable monthly, unlinked to headcount. At $7,205 PEPM the top percentile sits ~214× the panel's high-intensity cohort mean ($33.67); the median AI spender, at $12.50, has now passed 4.5× the low-intensity mean ($2.78) that defined the panel's null group. If the intensity gate is a threshold in PEPM rather than a rank, the share of firms above it rises every month this series does — the version of the question the monthly index can settle and the panel cannot.

A different gate on the same outcome: growth, not intensity (ICONIQ, September 2026)#

This page's result is intensity-gated: firms that spend heavily on AI vendors grow headcount, firms that dabble do not. ICONIQ's 2026 State of Scaling (ICONIQ Venture & Growth, September 2026, empirical — quarterly operating data from 137 software companies) measures the same outcome with a different gate on the left-hand side: revenue growth. Median headcount change by YoY revenue-growth cohort (n = 390 / 200 / 197 / 76 company-quarters):

YoY revenue growth	2022–23	2024	2025	2026
100%+	119%	65%	115%	146%
50–100%	47%	38%	46%	34%
25–50%	17%	12%	18%	6%
<25%	(6%)	(5%)	4%	2%

The shapes match, and that is the point. Both instruments produce a steep gradient in which the top of the distribution expands hard and the bottom is flat — this page's high-intensity adopters at ~10% headcount growth, ICONIQ's 100%+ growers at 146%. Neither is a substitution result. But the gates are not interchangeable and the difference is informative: AI-spend intensity and revenue growth are correlated but not the same variable, and ICONIQ never cuts this chart by AI status, so its version cannot distinguish "AI-forward companies hire" from "fast-growing companies hire," which is the confound this page's paper controls for and ICONIQ does not. Read as corroboration of direction on a second population (venture-backed software rather than the paper's economy-wide panel), not as a replication.

The population boundary, drawn inside one document. The same report collects public-company headcount reductions explicitly attributed to AI over 2026 — Amazon 9%, Dell 10%, Atlassian 10%, Snap 16%, Coinbase 14%, Cisco 5%, PayPal 20%, Intuit 17%, Cloudflare 20%, Meta 10%, GitLab 14%, Oracle 13%, monday.com 20%, Microsoft 2.1%, summarised as "roughly 10–20% this year" (secondary: a TechCrunch roundup of filings, memos and press coverage, with two entries flagged as outside estimates). So in the same twelve months, on the same publisher's evidence, private AI-forward hypergrowth added headcount at 146% and mature public incumbents cut 10–20%. That is this page's adopter-versus-native boundary showing up as two opposite-signed numbers in one report, and it is the cleanest available answer to why "does AI reduce headcount" keeps getting contradictory answers — AI Investment Story, Not Efficiency Story carries the revenue-side twin.

The mechanism ICONIQ observes, which is the thing this page's first open question asks for. Under "Perspectives from the ICONIQ Network," the transmission is not restructuring but backfill suppression: attrition "increasingly being used as an opportunity to avoid, delay, or down-level backfills"; "no-backfill policies… explicitly tied to AI-driven productivity improvements in sales, content, operational, and analytical functions"; "hiring freezes… used while companies evaluate AI-driven productivity gains before approving incremental headcount." Unquantified — an investor's qualitative read of its own portfolio, with no share of companies attached — but it names a net-headcount channel that operates without a single layoff, which is exactly the kind of mechanism a spend-and-headcount panel would see as a flat line and never explain.

What this panel settles about a formal substitution model#

Banerjee & Singh's HAT model (arXiv 2607.20781, July 2026, practitioner-opinion — a formal model with no data) derives seven testable predictions about when organizations replace workers with AI. This panel is the vault's only firm-level measurement of the outcome those predictions describe, and it bears on three of them:

• On P1 (discontinuous automation). The model predicts workforce change arrives as sharp phase transitions at threshold crossings — "spikes in automation-related layoffs or role reclassifications," not smooth diffusion. The high-intensity event study is the opposite shape: 0.003 → 0.020 → 0.071 → 0.188 → 0.277 → 0.452 at months 0/3/6/12/18/24, a compounding ramp on a learning curve. Adoption itself can look abrupt (OpenAI's internal functions went 20%→75% Codex token share in a single month); the workforce adjustment P1 actually names does not.
• On P2 (middle-management vulnerability). The model predicts AI thins middle layers before the top or bottom, automating coordination roles first. Here, manager-plus headcount grew +6.5% at high-intensity adopters while entry-level grew fastest at +12.0%; only the share moved the predicted way (manager-plus −1.52pp, entry-level +1.15pp). The compositional tilt is consistent with the direction; the levels are not, and the layer that shrank in share is not the layer that shrank in count. The counter-measurement runs the same way — Brynjolfsson/Chandar/Chen find 22–25-year-olds in the most AI-exposed occupations at −16% while 35–49-year-olds in those same occupations grew +8%.
• On P6 (deployment-scale acceleration). The model's Min$' = T_k/n_k + M'_k predicts larger firms cross substitution thresholds first, purely by amortizing training cost over more deployments. The selection table is consistent — adopters here are larger, more technical, higher-paying and more VC-backed before adoption — but confounded, and the variable that actually separated outcomes in this panel was per-employee intensity ($2.78 vs $33.67/month) and elapsed time, not headcount to amortize over. A firm can deploy across every seat and still land in the null group. That is the complements story, not the amortization story, and HAT's own limitation 2 concedes n_k is exogenous to it.

Expectation versus realization, graded by the same instrument (McKinsey, August 2026)#

This panel measures headcount from payroll-adjacent records. The other half of the workforce literature is survey expectation, and The state of AI in 2026 (McKinsey / QuantumBlack, 2026-08-25, empirical but self-reported) is the first source in the corpus that lets a forward workforce expectation be graded against the same instrument's later realization (Exhibit 16; 2025 n=1,753, 2026 n=1,521):

	Total decrease	Little or no change	Total increase	Don't know
2025 expectation for the coming year	32	43	13	12
2026 report of what happened over that year	14	66	8	13
2026 expectation for the coming year	39	43	10	8

The expectation overshot the realization by more than 2×. Two-thirds of respondents report little or no AI-related change in total employment over the year in which a third of them had expected reductions. The same pattern repeats at function level (Exhibit 17): in every one of eleven functions the reported decrease is smaller than the prior year's expected decrease, and in marketing and sales (33 expected → 14 reported), manufacturing (30 → 13), knowledge management (27 → 11) and strategy and corporate finance (28 → 15) it is roughly half. Software engineering is among the functions where realization came closest to the forecast (31 expected → 22 reported, a 0.71 ratio against 0.41–0.43 for knowledge management, marketing and sales, and manufacturing), and it carries the third-largest expectation for the year ahead (36, behind service operations at 44 and supply chain management at 38).

What this does and does not say about the result above. It is not a contradiction and it is not a supersede: the modal outcome in both instruments is a workforce that did not shrink because of AI, and this panel's positive headcount effect is gated on spend intensity that a global survey does not measure at all. The units differ in a way worth keeping straight — Revelio measures total headcount change at firms whose AI spend is observed, while McKinsey asks a respondent to attribute employment change to AI, which is a counterfactual judgment rather than a count, and the 13% "don't know" is honest about it.

The transferable lesson is calibration. The only occasion the corpus can grade a survey's forward AI-workforce expectation, it overshot by a factor of 2.3. That is the discount to carry into the 39% now expecting declines in the coming year — including where that expectation is quoted as evidence of labour-market effect. Caveats on the grade itself: the two waves are different respondent draws from the same panel rather than a tracked cohort, the 2026 wave asks a slightly different question (attributed change, not expectation), and a respondent who over-predicted has an incentive to under-report having done so.

Connections#

• Seniority-Biased AI Adoption: The Junior Share at Adopting Firms — the instrumented 41-country adoption design that finds the opposite composition tilt on the same Revelio seniority coding: junior share −1.9pp at adopters, driven by senior growth. On composition it is the better-identified source; this panel's +12.0% stands for US intensive spenders only

• The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay — the price side of the Indeed postings series: the exposure pay premium is mostly the same senior-tilted composition this page contrasts with Ramp's entry-level growth

• The Enterprise AI Adoption Gradient — the same object seen from the vendor's ledger instead of the customer's card: OpenAI's ChatGPT Enterprise account records linked to Compustat. It agrees with this panel that adoption is gated on something beyond firm size (there, FY2021 SG&A / R&D / capitalized-software stocks), and it is the one adoption instrument in the vault that cannot supply a rate — its public-company adopter set is a disclosure-motivated random sample, so its 417-adopter: 11,784-non-adopter counts are a sampling artifact and its adopter/non-adopter contrasts are attenuated lower bounds

• AI Product Economics Maturation — the same trade seen from inside one P&L: a $500M+ ARR fintech CFO describes token spend reaching "~5-10% of payroll" and then being pulled "out of the future headcount plan," i.e. AI spend substituting for hiring at the budget line rather than through a layoff — the mechanism this page's panel would record only as a flat headcount curve

• Organizational Complements to AI — the HAT substitution model above lives there in full, with its seven predictions checked against the rest of the vault; this page supplies the firm-level counter-evidence on P1, P2 and P6

• Task Saturation: Broad but Shallow AI Diffusion — the task-level shape consistent with headcount growth rather than displacement: AI reaching most occupations but a fifth of their tasks, with end-to-end automation the intent of under 10% of non-routine-cognitive conversations

• Systems Thinking Over Specialization — a named-firm corroboration of the entry-level-growth side: Netflix (an intensive AI adopter) calls intern/new-grad hiring "really important to our talent strategy," explicitly against the junior-decline narrative

• Organizational Complements to AI — the mechanism this page instantiates on payment traces: the intensity threshold (chat subscriptions do nothing; sustained coding-agent/API spend does) and the 6–12-month learning curve are the general-purpose-technology "complements lag" measured at the firm level — value arrives only after the complementary workflow/skill/org redesign, not at purchase

• AI Investment Story, Not Efficiency Story — independent corroboration of "investment, not efficiency" from a different dataset: adopting firms staff up (broadly, incl. entry-level and sales), i.e. hire ahead of the output, rather than shrinking — the headcount-side companion to Emergence's lower-revenue-per-employee-now finding

• Cost-per-Task Over Cost-per-Token — where the index's price and token-tier series are read as a verdict on model selection: the blended effective price per million tokens down 41% to $0.68 from a $1.15 March peak, and a buying population defaulting away from the strongest model on explicitly price-per-token grounds

• Standardize the Infrastructure, Not the Tools — where the September index's token-share cut lands as an organizational finding rather than a market one: firms imposing company-wide model defaults are exercising the portability the gateway argument promises, observed as an outcome in this instrument and with the mechanism invisible to it

• Telemetry vs. Survey Measurement — a third measurement instrument: revealed AI-vendor spend linked to workforce records, which the paper explicitly positions to "replace messy surveys and exposure measurements" (its Figure 1 shows survey adoption estimates ranging 18%→78% for the same period) — behavior-not-feeling, like Faros's telemetry, but on the adoption + labor margins

• Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — the axis this paper adds: the taxonomy's four measures are all occupation-level exposure; Ramp's is firm-level observed adoption (who actually paid, when, how much) — the firm-vs-occupation variation exposure indices structurally cannot capture (two firms with identical workers can differ sharply in adoption), a new measurement axis alongside the taxonomy's occupation-level and market-implied ones

• The Automation–Optimism Link — objective counter-evidence to the entry-level/junior job-loss fear recorded there as perception: at intensive adopters, entry-level headcount grew +12% and its share rose

• Founder as Agent Orchestrator — a counter-signal to the lean-team reading, with a population caveat: these are established US firms adopting AI and growing ~10%, not AI-native startups built lean from day one — AI-at-adopting-firms adds headcount rather than substituting for it (consistent with the "average AI company staffs up" note on AI Investment Story, Not Efficiency Story)

• Returns to Expertise in Agentic Coding — the labor-demand reading of the expertise premium: the Indeed rebound is 71% senior and 37% AI-titled, i.e. employers posting for experienced professionals who can work with AI. Where that page measures the premium inside sessions, this is the same premium priced in vacancies — and it is the compositional evidence that most cuts against this page's entry-level result

• The Tragedy of the Cognitive Commons — the pipeline question the postings data speaks to directly: postings did recover in the most-exposed occupation, and the recovery was senior. So a vacancy rebound is not by itself evidence the regeneration mechanism came back, which is exactly the distinction that page draws between recovering positions and recovering developmental content

• AI-Native Startup Lifecycle — same counter-signal at the lifecycle level: the "headcount stays flat through Scale" thesis describes AI-native firms; this paper's adopters are incumbents whose intensive AI use coincides with expansion, so "AI ⇒ leaner" is population-specific, not universal

• The Open-Weight Frontier Gap — the demand-side half of the same monthly index: the highest-PEPM tail of this page's intensity distribution ($248/employee/month, ~7× the high-intensity cohort mean) is exactly the group buying model-serving platforms, and it pays the American labs at above base rates — so open/Chinese-model adoption in this instrument is additive to frontier-lab spend, not a substitution for it

• Acceleration Whiplash — a third spend instrument, and the first that reports spend next to engineering outcomes. DX's Q2 2026 panel (vendor-claim, 500+ customer organizations) puts median quarterly organizational AI spend at ~$1.5K rising to ~$44K over four quarters, with tech-sector spend up nearly 28x — a different denominator from this page's two (Ramp's PEPM is per-employee-per-month; DX's is per-org-per-quarter), so the series are not comparable without a headcount assumption. One is available as a rough consistency check rather than a finding: ~$44K/quarter is ~$14.7K/month, which at the Ramp AI Index's June-2026 median of $10.59 per employee per month implies an organization of roughly 1,400 employees — plausible for a panel of enterprises buying a developer-productivity platform, and a reminder that DX's median org is much larger than Ramp's. (This vault's arithmetic on two vendors' medians, not either vendor's claim.) What DX adds beyond a third series is the outcome side: it reports the spend climbing while its own innovation ratio stays flat, i.e. spend accelerating faster than any measured conversion — the engineering-metrics counterpart to this page's finding that intensity buys headcount, with the mechanism still unnamed on both sides

• AI and Market Power — the competition-outcome sibling on European administrative microdata, and the source that most needs this page's intensity gate to be read correctly: OECD find AI-using French and Portuguese firms gaining no market-share rank and no markup growth over five years, but their adoption measure is a binary ICT-survey question that cannot separate the $2.78-PEPM margin from the $33.67 one. Two independent corroborations run the other way — both papers localise the measurable effect to the same sector (Information here, ICT there), and OECD's own future-work list asks for exactly the intensity measure this panel supplies

• Build Instead of Buy Under Agentic Coding — the same payment-rail logic pointed at a different vendor category. Its second open question asks for this page's instrument applied to software spend rather than AI spend, because a genuine build-instead-of-buy displacement should show up as decelerating net-new SaaS vendor spend in the industries reporting it

• Balance-of-Power Superintelligence — the prediction this panel is the measured counterpart to. Zuckerberg's manifesto claims "recent statistics suggest" individual capability growth may outpace automation, cites none, and predicts more firms with fewer people each; the intensity gate found here says the employment effect is real but conditional on adoption depth, not a broad tide

Open Questions#

• What operational mechanism converts intensive AI spend into hiring? The paper establishes the correlation (adopters, especially intensive ones, grow) but explicitly cannot say why — product acceleration, sales productivity, engineering leverage, support automation, faster analysis, or new business lines are all candidates, and the firms that cracked it have no incentive to share. Related evidence (2026-08-04): the monthly AI Index narrows what the top of the intensity distribution is buying — the $248-PEPM cohort is defined by paying model-serving and inference platforms, i.e. building on APIs rather than buying more seats. That is a characterisation of the spend, not of the mechanism, but it points the candidate list toward engineering/product leverage and away from enterprise chat rollout. A named channel, from inside the firms (2026-09-22): 2026 State of Scaling: The Great Sorting observes ICONIQ's portfolio running backfill suppression rather than restructuring — attrition used to 'avoid, delay, or down-level backfills,' explicit no-backfill policies 'tied to AI-driven productivity improvements in sales, content, operational, and analytical functions,' and hiring freezes held while productivity gains are evaluated before incremental headcount is approved. This is the first mechanism in the corpus that would be invisible to this page's instrument: it changes net headcount with no layoff event and no hiring event, so a spend-and-headcount panel records only a flat line. It is also entirely unquantified — an investor's qualitative observation of its own book, no share of companies, no effect size — and it is a mechanism for headcount not growing, whereas this bullet asks what converts spend into growth. Worth carrying as a candidate for the other tail of the distribution.
• Does the effect diffuse beyond Information as adoption cohorts mature? Significant gains are, so far, an Information-sector phenomenon; the authors intend to update with later cohorts and post-24-month windows. Will professional services, finance, and non-technical sectors follow, or is the coding-agent workflow special? Independent corroboration of the sector boundary (2026-08-11): OECD AI Papers No. 62 finds the markup premium from AI patenting is significant only in ICT (AI × ICT +7.95%, while standalone AI turns negative with fixed effects) across ~600K firm-years in 21 European countries — a different outcome (markups, not headcount), a different instrument (patents, not spend), a different continent, and the same sector line. Their reading is the sharper version of the question: AI pays where it is the firm's output, not where it is an input. Not an answer to diffusion-over-time, but two instruments now agree on where the effect currently lives.
• Is the entry-level growth durable or a lead-indicator that later reverses? Gains compound through month 24 on thinning samples; whether the +12% entry-level result holds (or inverts toward the Brynjolfsson "Canaries" pattern) as high-intensity adopters mature past 24 months is unresolved. Countervailing signal (2026-08-04): Indeed Hiring Lab finds the May 2025 – May 2026 software-postings rebound is 71% senior roles, on a later window than this panel's average and on the demand flow rather than the headcount stock — not an answer (different unit, different population, no control group), but the first vault evidence pointing the other way on composition. Partially answered (2026-09-22): Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (empirical, ADP payroll microdata through June 2026) settles the Canaries half of the comparison — the pattern this question asks whether the +12% might invert toward has itself continued and widened, to a 19% kept-pace shortfall for 22–25-year-olds in exposed occupations, driven by hiring rather than separations, and concentrated where AI usage is automative. It does not settle the durability question, and the reason is structural rather than a data gap: ADP firm identifiers are anonymized, so Canaries cannot observe adoption at the firm level at all and cannot say what happened inside intensive adopters past month 24. The two instruments measure different objects and are reconciled rather than averaged (see the section above). Contradicted on composition (2026-10-01), not on durability: Seniority-Biased AI Adoption: The Junior Share at Adopting Firms finds adopters' junior share falling through March 2026 on the same Revelio coding, and falling further with adoption intensity. The one overlapping cut — state-level adoption tiers from Anthropic Economic Index usage — leans against this panel: the young-worker decline is sharpest in leading-adoption states (about −19% for the most-exposed quintile) and smallest in emerging-adoption ones. Geography is a weak proxy for firm adoption, so this sharpens the question rather than answering it: the durable test remains a firm-level panel that can follow the same adopters past 24 months. A second Indeed series leans the same way (2026-10-01): AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It reports the entry-level share of salaried postings in the most-exposed occupations falling 29% → 10% (2021 → 2026) — economy-wide flow, not adopter stock, so still not the test; detail on The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay.
• Is the top-1% per-employee-spend decline a seasonal dip, a vintage artifact, or the start of a plateau in the intensity distribution's top tail? The three readings separate in the data: a seasonal dip reverses by October (Ramp names August engineer holidays and observes the same dip each November–December), a vintage artifact shrinks once August is revised on the same footing as July — this page's cross-edition arithmetic puts the like-for-like fall at roughly 2.6% against a published 9.7% — and a genuine plateau shows up as a second consecutive decline in an already-revised series. Trigger: the October and November 2026 editions of the Ramp AI Index, which republish the same cohort series.

Sources#

• How Does AI Change Labor Demand? Evidence from 41 Countries — Chandar & Klein Teeselink, How Does AI Change Labor Demand? Evidence from 41 Countries (Stanford Digital Economy Lab working paper, 2026-09-20, 156pp, empirical). Cited only for §5.1 Table 3 (junior share, junior/senior/total employment), §5.3.2 (matched-design pre-trends, intensity measure), and its shared Revelio seniority 1–2 coding. Full treatment and parse warnings on Seniority-Biased AI Adoption: The Junior Share at Adopting Firms
• AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It — Jack Kennedy, Indeed Hiring Lab, 2026-09-17 (empirical, vendor postings data, same GSTI exposure instrument as Gallacher). Cited only for the entry/senior posting-share shift; full treatment on The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay
• A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment — Kharazian, Simon & Stevens, A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment (Ramp Economics Lab × Revelio Labs, June 30 2026): §3 Data (Ramp AI Index, adoption/PEPM definitions, Revelio seniority), §4 selection + adoption gradients, §5 Callaway–Sant'Anna methodology, §6 Results (Tables 3–6: headcount, shares, sector, never-treated), §7 Conclusion
• AI and Job Postings: From Destruction to Creation? — Guillermo Gallacher, AI and Job Postings: From Destruction to Creation? (Indeed Hiring Lab, 2026-07-08; empirical, ~1,158-word blog post). Key points list and §"A senior, AI-fluent rebound" for the 15% / 7% / 27.5% / 71% / 37% figures; §"Are other occupations exposed to AI experiencing a similar rebound?" for the 2022–2026 vs 2025–2026 sign flip; conclusion for the seniority-biased-technological-change concession. Conflict of interest: Indeed's research arm analyzing Indeed's own postings data — one job board's vacancy flow, not a labor-market census. Not cited, because the numbers exist only inside chart images hosted on the source site with no local copies: the identities and per-country shares of the six economies in the international chart (the prose names only Germany and France as exceptions), and both scatter plots' correlation coefficients and p-values (the prose asserts significance without reporting either). No causal estimate is available from this source — the design has no control group and no exposure variation beyond the occupational cross-section.
• Ramp's latest data on China vs. the American AI Labs — Ara Kharazian, Ramp's latest data on China vs. the American AI Labs (Ramp AI Index monthly letter, 2026-07-08; empirical, ~840 words of prose). The prose supplies 42.4%/39.5%, 5.8% up from 4.5%, $248 vs $10.59, and 85.8%/93.2%/96.4%. Everything else quoted here comes from the recovered chart datasets: all four in-article charts were Datawrapper iframes with no static fallback, and the ingest pass pulled each chart's own dataset.csv and reproduced the full monthly series in the raw file (vendor share from 2023-01, the other three from 2023-07). Those tables are the only source for the crossover date, OpenAI's November-2025 peak, and the Google/Microsoft/xAI/DeepSeek lines — the letter names none of those four vendors. Conflict of interest and sampling: Ramp measures its own corporate-card/bill-pay customer base (VC-forward-skewed, not a random sample of US businesses) and sells itself as the authoritative AI-adoption data source; the raw doc's note: field records both. Derived-not-stated figures are labelled inline: the AI-spender-rebased penetration series, the 82.5%-use-both / 3.6%-use-neither decomposition, the 23.5×-and-compressing multiple, and the comparison of $248 PEPM against the paper's $33.67 high-intensity mean are all this vault's arithmetic on Ramp's tables, not Ramp's claims. Method and denominators are documented only at ramp.com/data/ai-index, which is not in the corpus
• China, Open Source & AI Competitiveness — Andrew Ng — Andrew Ng interviewed by James Hohmann, Washington Post Live "Building America" (2026-07-29, 30:39; practitioner-opinion). Used only for the "third voice" paragraph above: the postings claim, the no-job-apocalypse position, the complementarity mechanism, and the software-as-harbinger generalization. No data is offered for any of it and none of the quantities are sourced; the software-to-other-sectors extension is a prediction. Transcript is YouTube auto-captions with ASR proper-noun corrections applied at ingest — the labor passages are intact, though the caption turn markers make speaker attribution reliable in substance rather than verbatim
• The Human-AI Substitution Principle: When will you be replaced by AI in your organization? — Banerjee & Singh, arXiv 2607.20781 (2026-07-22; practitioner-opinion, formal model, no empirical data): §5.6 Table 3 (predictions P1, P2, P6 and their stated empirical signatures), §4.5.2 Theorem 14 + Corollary 6 (the discontinuity claim), §4.3 Corollary 2 (middle-management vulnerability, conditional on a single-crossing assumption the paper hand-sets), §5.6.6 + §5.8 limitation 2 (deployment-scale amortization with n_k exogenous). Full treatment at Organizational Complements to AI
• Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think — Andrew Ng interviewed by Marina Mogilko, Silicon Valley Girl (2026-08-28, practitioner-opinion): the postings claim repeated, the busier-than-ever and can't-find-skilled-people assertions, the AI-native interns, and the reskilling condition. No data offered
• The state of AI in 2026: On the road to ROI — Dan Tinkoff, Lieven Van der Veken & Michael Chui with Tara Balakrishnan, The state of AI in 2026: On the road to ROI (McKinsey / QuantumBlack, 2026-08-25, empirical, self-reported; online survey, 1,719 participants in 97 nations, fielded May 4 - June 8 2026, GDP-weighted). Cited here for Exhibit 16 (the 2025-expectation / 2026-realization / 2026-expectation triptych), Exhibit 17 (the same at function level) and the key-takeaways framing. Self-report, one respondent per organization, attribution to AI left to the respondent; COI — McKinsey sells AI transformation consulting.
• September 2026 Ramp AI Index: Cracks in the AI thesis, part 2 — Ara Kharazian, September 2026 Ramp AI Index: Cracks in the AI thesis, part 2 (2026-09-09; empirical, ~3,600 characters of prose over the same card/bill-pay rail). The prose supplies 56%/+0.4pp, 43.8%/39.8%, the −9.7% top-1% decline and its $7,976 → $7,205 endpoints, $0.68 per million tokens and the 41% fall from a $1.15 March peak, 45% frontier token share, and the 6.4%/3.6% open-serving figures. Everything else quoted here comes from the five recovered Datawrapper datasets (chart ids HjOZl, kBIUV, 0lGNa, W0I7q, ibX1m; the page has no static chart images, so the ingest pass pulled each chart's own dataset.csv): the monthly increments, the xAI/Microsoft/Google/DeepSeek lines, the three-cohort PEPM series, and the cross-edition revision comparison. Two provenance traps, both disclosed in the raw's note: and both load-bearing: the spend CSV's July top-1% cell is the revised $8,024.58 while the prose quotes the as-reported $7,976, and the token-price CSV's tail (0.603 on 2026-09-02) trails the prose's "$0.68 as of this week" because the CDN-cached endpoint lags the live client-rendered chart — prose is authoritative for current values, the CSVs for the shape of the series. COI unchanged from the July row: Ramp measures its own VC-forward-skewed customer base and markets itself as the authoritative AI-adoption dataset. Instrument provenance at Ramp
• 2026 State of Scaling: The Great Sorting — ICONIQ Venture & Growth, 2026 State of Scaling: The Great Sorting (September 2026, empirical; 52-page PDF, docling-parsed; quarterly operating data from 137 software companies, portfolio plus 11 selected publics). Cited here for the headcount-change-by-revenue-growth-cohort series (p.43, read off the page image in a two-pass and cross-checked against pdftotext -layout), the public-company workforce-reduction timeline (p.20 — secondary, a TechCrunch roundup of SEC filings, executive memos and press coverage, hand-transcribed at ingest from an image-only chart) and the ICONIQ-network backfill observations (p.43). Two caveats: n is company-quarters, not companies (2026 = 76, down from 197), and the chart is cut by revenue growth and never by AI status, so it cannot separate an AI effect from a growth effect. Publisher COI at ICONIQ
• Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence — Brynjolfsson, Chandar & Chen, Canaries in the Coal Mine?, Stanford Digital Economy Lab, revised August 2026, 140pp (empirical, ADP administrative payroll microdata through June 2026). The counter-instrument, now read firsthand rather than through secondhand citation. Cited here for §1 (the 13%/16% regression vintages superseded by the 15%/19% descriptive measure), §2.2 (the ~11% / ~10% quintile levels), §2.4 (hiring rather than separations), §6.3 (state-level adoption tiers, the only cut overlapping this panel), §1.1 and §6.3 (anonymized firm identifiers, the structural reason the two instruments cannot be joined) and Appendix L (the within-firm Poisson estimate attenuating across vintages). Parse notes: benign table-collapse on repeated panel headers (verified against pdftotext -layout); Table 2 is row-shifted and is cited nowhere; Table A.6 is welded past rank 25. Full ledger in Source Notes